Abstract
There are explicit tradeoffs between generality and specificity in computer vision methods with respect to designing impactful end-to-end solutions for ecological challenges — computer vision systems that are general purpose but optimal for each stakeholder, taking into account cost, human effort, and risk. We explore these tradeoffs across several dimensions, investigating the impact of generality vs specificity with regards to data, labels, tasks, and models. We investigate how various methods that attempt to achieve the best of both worlds perform across increasingly large shifts in real-world ecological benchmarks across data modalities, including task-specific dataset subselection, domain adaptation, expert-in-the-loop retrieval systems, and task-aware compression.